Jawad00100/C-to-Cuda

CUDA to C Code Conversion Dataset

A large-scale dataset containing 8257 pairs of CUDA code and their equivalent C code conversions, generated using Google's Gemini 2.5 Flash model.

Dataset Description

This dataset provides high-quality CUDA to C code translations for training code translation models, studying parallel programming patterns, and developing automated code conversion tools.

Source Data

• —Original Dataset: mzubai/cass - CUDA code samples
• —Conversion Model: Google Gemini 2.5 Flash with structured output
• —Processing: Parallel processing with 10 API keys, rate limiting, and quality filtering

Dataset Statistics

• —Total Samples: 8257 CUDA-C pairs
• —Format: CSV with structured columns

Dataset Schema

Usage Examples

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("your-username/cuda-to-c-conversion")

# Access the data
train_data = dataset['train']
print(f"Dataset size: {len(train_data)}")

# Example sample
sample = train_data[0]
print("CUDA Code:", sample['cuda_source'][:200] + "...")
print("C Code:", sample['c_code'][:200] + "...")

Training Code Translation Models

# Prepare for sequence-to-sequence training
cuda_codes = [sample['cuda_source'] for sample in train_data]
c_codes = [sample['c_code'] for sample in train_data]

# Use with transformers
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("codet5-base")
model = AutoModelForSeq2SeqLM.from_pretrained("codet5-base")

# Tokenize inputs
inputs = tokenizer(cuda_codes, truncation=True, padding=True, return_tensors="pt")
targets = tokenizer(c_codes, truncation=True, padding=True, return_tensors="pt")

Data Analysis

import pandas as pd

# Convert to pandas for analysis
df = pd.DataFrame(train_data)

# Basic statistics
print(f"Average CUDA code length: {df['cuda_source'].str.len().mean():.0f} chars")
print(f"Average C code length: {df['c_code'].str.len().mean():.0f} chars")

# Code complexity analysis
cuda_lines = df['cuda_source'].str.count('\n').mean()
c_lines = df['c_code'].str.count('\n').mean()
print(f"Average lines - CUDA: {cuda_lines:.1f}, C: {c_lines:.1f}")

Dataset Creation Process

1. 1.Source: Started with CUDA code from mzubai/cass dataset
2. 2.Conversion: Used Google Gemini 2.5 Flash with structured JSON output
3. 3.Parallel Processing: 10 API keys with round-robin distribution and rate limiting
4. 4.Quality Control: Removed failed conversions and error responses
5. 5.Validation: Structured output validation using Pydantic models

Quality Assurance

Conversion Quality

• —Structured Output: JSON schema validation ensures consistent format
• —Error Filtering: Only successful conversions included (9.4% error rate removed)
• —Manual Validation: Sample verification of conversion accuracy
• —No Artifacts: Removed API error messages and malformed responses

Data Integrity

• —Unique IDs: Each sample has a unique identifier
• —Complete Pairs: All samples have both CUDA and C code
• —Size Limits: Handles large code files (up to 1MB per field)
• —Encoding: UTF-8 encoding for international character support

Use Cases

Research Applications

• —Code Translation Models: Train CUDA↔C translation systems
• —Programming Language Analysis: Study syntax and semantic differences
• —Parallel Programming: Analyze parallelization patterns
• —Code Generation: Develop automated code conversion tools

Educational Applications

• —Learning Resources: Examples of CUDA to C conversions
• —Curriculum Development: Teaching parallel programming concepts
• —Code Comparison: Understanding different programming paradigms

Industrial Applications

• —Legacy Code Migration: Automated CUDA to C conversion
• —Performance Analysis: Compare parallel vs sequential implementations
• —Code Optimization: Study efficient coding patterns

Limitations and Considerations

Technical Limitations

• —Semantic Accuracy: Conversions may not be functionally identical
• —Compilation: Generated C code may require manual adjustments
• —Context: Limited by model's context window for very large files
• —Complexity: Very complex CUDA kernels may not convert perfectly

Usage Guidelines

• —Validation Recommended: Test converted code before production use
• —Manual Review: Complex conversions should be manually verified
• —Compilation Testing: Verify that generated C code compiles
• —Performance: Converted code may have different performance characteristics

Dataset Metrics

Citation

If you use this dataset in your research, please cite:

@dataset{cuda_to_c_conversion_2024,
  title={CUDA to C Code Conversion Dataset},
  author={Generated using Google Gemini 2.5 Flash},
  year={2024},
  url={https://huggingface.co/datasets/your-username/cuda-to-c-conversion},
  note={Derived from mzubai/cass CUDA dataset},
  samples={2043}
}

License

This dataset is released under the same license terms as the original mzubai/cass dataset. Please refer to the original dataset for specific license information.

Acknowledgments

• —Original Dataset: mzubai/cass for providing the CUDA source code
• —Conversion Model: Google Gemini 2.5 Flash for high-quality code translation
• —Processing: Parallel processing pipeline for efficient dataset generation

Updates and Versions

• —v1.0: Initial release with 8257 CUDA-C pairs
• —Quality: Error-filtered, validated conversions only
• —Format: CSV with id, filename, cudasource, ccode columns

Dataset Size: 2043 samples File Format: CSV Encoding: UTF-8 Quality: Production-ready, error-filtered